Classification of Multiple Steganographic Algorithms Using Hierarchical CNNs and ResNets

In general, image deformations caused by different steganographic algorithms are extremely small and of high similarity. Therefore, detecting and identifying multiple steganographic algorithms are not easy. Although recent steganalytic methods using deep learning showed highly improved detection accuracy, they were dedicated to binary classification, i.e., classifying between cover images and their stego images generated by a specific steganographic algorithm. In this paper, we aim at achieving quinary classification, i.e., detecting (=classifying between stego and cover images) and identifying four spatial steganographic algorithms (LSB, PVD, WOW, and S-UNIWARD), and propose to use a hierarchical structure of convolutional neural networks (CNN) and residual neural networks (ResNet). Experimental results show that the proposed method can improve the classification accuracy by 17.71% compared to the method that uses a single CNN.